数字化外科医生：大数据自动化和人工智能外文文献

The digital surgeon How big data automation and artifi cial James Wall Thomas Krummel Division of Pediatric Surgery Stanford University Stanford Byers Center for Biodesign Stanford University a b s t r a c ta r t i c l ei n f o Article history Received 29 August 2019 Accepted 19 September 2019 Key words Future pediatric surgery Automation and artifi cial intelligence in pediatric surgery Exponential growth in computing power data storage and sensing technology has led to a world in which we canbothcaptureandanalyzeincredibleamountsofdata Theevolutionofmachinelearninghasfurtheradvanced the ability of computers to develop insights from massive data sets that are beyond the capacity of human anal ysis The convergence of computational power data storage connectivity and Artifi cial Intelligence AI has led to health technologies that to date have focused on diagnostic areas such as radiology and pathology The ques tion remains how the digital revolution will translate in the realm of surgery There are three main areas where theauthorsbelievethatAIcouldimpactsurgeryinthenearfuture enhancementoftrainingmodalities cognitive enhancement of the surgeon and procedural automation While the promise of Big Data AI and Automation is high there have been unanticipated missteps in the use of such technologies that are worth considering as we evaluate how such technologies could should be adopted in surgical practice Surgeons must be prepared to adopt smarter training modalities supervise the learning of machines that can enhance cognitive function and ultimately oversee autonomous surgery without allowing for a decay in the surgeon s operating skills 2019 Elsevier Inc All rights reserved Contents 1 Current applications of AI in medicine S48 2 Enhancing current training modalities S48 3 Cognitive enhancement S48 4 Automation S48 5 Cautionary tales from other sectors S49 6 Conclusions S49 7 Next steps for pediatric surgeons S49 References S49 Ithasbeensaidthat DataisthenewOil inreferencetoashiftinthe world s most valuable resource 1 Exponential growth in computing power data storage and sensing technology has led to a world in which we can both capture and analyze incredible amounts of data The evolution of Machine Learning ML has further advanced the abil ity of computers to develop insights from massive data sets that are be yond the capacity of human analysis Specifi cally within medicine the near ubiquitous adoption of the Electronic Medical Record EMR has exponentially expanded the amount of medical data and has begun to reveal insights into practice patterns healthcare utilization and ulti mately patient outcomes The convergence of computational power data storage connectivity and Artifi cial Intelligence AI has led to health technologies that to date have focused on diagnostic areas such as radiology and pathology Many believe that the ultimate goal of these technologies is precision medicine that could achieve customi zation of healthcare with medical decisions treatments practices or products being tailored to the individual patient The question remains how the digital revolution will translate in the realm of surgery The birth of AI is generally credited to Alan Turing who famously broke the Enigma code of the German army that changed the course Journal of Pediatric Surgery 55 2020 S47 S50 Confl ictofInterestStatement Dr Wallhasnorelevantconfl ictstoreport Dr Krummel has ownership interest in Procept Biorobotics Corresponding author at Division of Pediatric Surgery Stanford University 300 Pas teur Drive Alway Building M116 Palo Alto CA 94305 Tel 1 650 723 6439 E mail address jkwall stanford edu J Wall https doi org 10 1016 j jpedsurg 2019 09 008 0022 3468 2019 Elsevier Inc All rights reserved Contents lists available at ScienceDirect Journal of Pediatric Surgery journal homepage intelligence will change surgical practice of WorldWarII Hesubsequentlyworkedonearlycomputinghardware and wrote the seminal paper entitled Computing machinery and intel ligence 2 He further proposed an experiment known as the Turing Test which proposed that a computer can be said to possess AI if it can mimic human responses under specifi c conditions More recently we have seen the realization of Big Data which can be defi ned as ex tremely large data sets that may be analyzed computationally to reveal patterns trends andassociations especiallyrelatingtohumanbehavior andinteractions WiththeavailabilityofBigData AIcanthenbeapplied to gain deeper insights into data AI can be defi ned as algorithms that give machines the ability to reason and perform cognitive functions such as problem solving object and word recognition and decision making We have recently seen advances in ML which is considered a branch of AI where algorithms enable computers to process data and learn on their own without constant supervision This is a signifi cant shift from the initial world of computer programing in which machines followed logical programs in which human operators had to anticipate and code for all possible situations In a recent example of ML success the AlphaZero algorithm devel oped by the DeepMind subsidiary of Google achieved a superhuman level of play in the games of chess and shogi Japanese chess and Go in less than 24 h The algorithm learned this level of skill by going through nearly infi nite iterations of self play without any prior domain knowledge other the rules of the game The superhuman capabilities of the algorithm were then proven by the system convincingly defeatinga world champion program in each case 3 1 Current applications of AI in medicine Radiology and Pathology have seen the highest surge in FDA ap proved AI systemsin recent years 4 5 The reason for thedevelopment ofAIinthesefi eldsislargelybecauseofthemanageabledigitalinputsof radiographs and pathology slides into ML algorithms It is also because ofthefactthatradiographicandpathologicalinterpretationdoesnotre quire real time interpretation Table 1 is a list of 14 FDA approved AI systemsthroughtheendof2018compiledbyDr Eric Topol 6 whore cently estimated that an additional 12 approvals have come in the fi rst half of 2019 These systems are all diagnostic in nature and rely on lim ited highly structured digital inputs such as EKG EEG DICOM radio graphic images retinal images etc These systems are analogous to facial recognition algorithms that rely on complex image analysis and ML to differentiate patterns Surgical operations are more complex than diagnostic specialties in that the operations require preoperative planning real time decision making to perform the operation based on highly complex inputs including anatomic position video instrument movement etc and postoperative analysis to improve the quality of future procedures As such surgery is more analogous to the development of the self driving car than to the development of facial recognition The self driving car must plan for the trip make real time decisions that drive the car based on an array of sensor data and then analyze performance to im prove future trips So how close are we to realizing AI driven autonomous surgical pro cedures Maybe not as far as we might think Autonomous car devel oper Waymo was given regulatory permission to operate fully autonomous cars in parts of California in April of 2018 7 This would argue that at an order of magnitude the computation ability to auto mate and apply AI to surgical procedures is possible There are three main areas where the authors believe that AI could impact surgery in the near future enhancement of training modalities cognitive enhancement of the surgeon and procedural Automation 2 Enhancing current training modalities One of the early targets for AI in surgery is enhancement of training modalities This is a relatively low risk environment to establish the value of AI to surgery The testing of sensor technology that enables movement tracking and localization of anatomic landmarks is well suited for optimization in the training environment A large volume of research has focused on movement tracking and data have shown that reducing surgical variation and ineffi ciencies can result in better outcomes ML is now being applied to rich data sets from training to stratify surgical skill and recommend personalized training strategies to improve individual defi ciencies 8 9 3 Cognitive enhancement The volume of medical literature is already beyond the capacity of the individual physician to maintain up to date knowledge and it continuestogrowatanacceleratingpace 10 Forsurgeons thereisan other dimension of operative experience that is limited to an individ ual s capacity to see and perform cases There is a collective experiential knowledge of all surgeons that could offer incredible in sights if accessible to any individual surgeon especially at a time of need such as encountering an unexpected fi nding during an operation Imagine if instead of calling a partner with more experience a surgeon could instantly tap into the collective experience of experts around the world Collective surgical consciousness may seem a lofty goal but it may not be as far away as we think ThecombinationofBigDataandAIfueledbyoutcomesfromEMRsis emerging as a validated tool to enhance the decision making ability of the physician through clinical decision support Initial clinical decision support focused on treatment decisions in areas such as cancer which require huge data sets of literature but do not need real time inputs IBM Watson has been shown to match recommendations of expert tumor boardsin breast cancer treatment that mightoffer smallerhospi tals access to clinical care recommendations that would not otherwise be possible 11 Clinical decision support that adds real time physio logic data to the literature has been shown to improve performance in the management of sepsis 12 Finally the building blocks of using col lective experience and AI for real time intraoperative decision support are being built by showing the ability to interpret streaming video feeds and develop algorithms for workfl ow analysis 13 14 4 Automation Surgery is watching the dawn of autonomy in surgical robotics and initial attempts to combine AI with autonomous procedures The Mako system Stryker Kalamazoo MI for total joint replacement uses sophisticated preoperative planning to enable autonomous robotic bone preparation for joint replacement The system has been shown Table 1 FDA approved AI systems through the end of 2018 adapted from Topol 2019 6 CompanyFDA Approval Date Clinical Indication AppleSep 18 Atrial fi brillation detection AidocAug 18CT brain bleed diagnosis iCADAug 18Breast density via mammography Zebra MedicalJul 18Coronary calcium scoring Bay LabsJun 18Echocardiogram EF determination Neural AnalyticsMay 18Device for paramedic stroke diagnosis IDxApr 18Diabetic retinopathy diagnosis IcometrixApr 18MRI brain interpretation ImagenMar 18X ray wrist fracture diagnosis Viz aiFeb 18CT stroke diagnosis ArterysFeb 18Liver and lung cancer MRI CT diagnosis MaxQ AIJan 18CT brain bleed diagnosis AlivecorNov 17 Atrial fi brillation detection via Apple Watch ArterysJan 17MRI heart interpretation S48J Wall T Krummel Journal of Pediatric Surgery 55 2020 S47 S50 in studies to decrease soft tissue damage during bone preparation com paredtoexperiencedsurgeons 15 andtodecreaserecoverytimes 16 A similar approach has been taken by the Aqublation system Procept Surgical RedwoodCity CA forthetreatmentofbenignprostatichyper trophy Thesystem usespreoperativeplanningandreal timeintraoper ative ultrasound to perform an automated prostatic ablation with a precise waterjet technology Initial trails show equivalent urinary fl ow with potential benefi ts to the precision approach in terms of functional outcomes 17 18 While these systems show the promise of precision with Automa tion and the ability to incorporate real time sensor data into procedural decisions neither of these technologies has yet to incorporate AI Early attempts at using AI for surgical Automation have focused on task de construction and autonomous performance of simple tasks such as su turing 19 Such efforts are critical to establishing a foundation of knowledge for more complex AI tasks A recent publication showed the ability of an AI driven robot to perform a superior bowel anastomo sis to expert surgeons in porcine tissue as measured by the consistency of suturing informed by the average suture spacing the pressure at which the anastomosis leaked the number of mistakes that required removing the needle from the tissue completion time and lumen reduction 20 5 Cautionary tales from other sectors While the promise of Big Data AI and Automation is high there have been unanticipated missteps in the use of such technologies that are worth considering as we evaluate how such technologies could should be adopted in surgical practice TranslationofAIandAutomationtodifferentenvironmentscanpose signifi cant challenges It is one thing for an autonomous car to learn to drive on the sunny streets of California but an entirely different chal lenge to translate that learning to driving in Minnesota in January with poor visibility and snow 21 The history of pediatric surgery has often seen our profession adapt adult devices for use in children We must consider the ramifi cations of this as surgical training enhanced cognitivecapabilities andultimatelyAutomationwilllikelybeginbyfo cusing on large data sets from adults ML must be supervised to ensure optimal outcomes While the human brain does not have the computational capacity to learning from massive data sets we still retain an advantage in making sense of the results During an early iteration in applying ML to the diagnosis of pneumonia the systems were given two sets of chest radiographs with and without pneumonia to learn the difference It happened that a majority of the fi lms positive for pneumonia were from one hospital and the majority of normal fi lms were from another The algorithm quickly determined that the most predictive feature of pneumonia was the fi ducial used to indicate right and left side on the radiograph which turned out to be different between the two hospitals 22 While computer scientists would argue that with more data the com puter would ultimately get this right this does highlight a dimension of reasoning that human intelligence can still contribute In pediatric surgery this is particularly important as we will likely have to deal with smaller data sets than adult surgeons and thus may require ongo ing human interpretation The loss of skills to Automation is a real issue that must be proac tively addressed While Automation has made the latest generations of commercial aircraft extremely safe and easy to fl y it comes with the risk of eroding the skills of pilots in unanticipated situations In the tragic case of Air France 447 the loss of the airspeed sensor caused the plane to switch off automated fl ights controls as the plane experienced turbulence The automated fl ight control which would normally pre vent the pilot from being able to lift the nose of the plane up enough to stall the aircraft disengaged without the real time sensor data needed for it to make decisions The pilot apparently unaware of losing automated fl ight control went on to pull the nose up to a point of stalling the aircraft with disastrous consequences 23 As surgeons we must understand the effects of Automation on our skill sets and de velop effective countermeasures to maintain patient safety in unantici pated situations 6 Conclusions Big Data AI and Automation will fundamentally change the practice of surgery in the future Surgeons must be prepared to adopt smarter training modalities supervise the learning of machines that can en hance cognitive function and ultimately oversee autonomous surgery without allowing for a decay in manual surgical skills 7 Next steps for pediatric surgeons In order to prepare for the coming evolution in surgery the authors offer several recommendations to consider as a society Prioritize the right clinical needs in Pediatric Surgery that could bene fi t from AI and Automation Join forces tocollect data in anoptimizedformat for machinelearning Oversee the learning and implementation of technology in our fi eld Track outcomes Develop countermeasures to ensure themaintenance of surgicalskills References 1 The world s most valuable resource is no longer oil but data Ecomonist https no longer oil but data Accessed 8 15 19 2 Carpenter BE Doran RW Turing AM Turing AM Woodger M A M Turing s ACE re port of 1946 and other papers Cambridge Mass Los Angeles MIT Press Tomash Publishers 1986 3 Silver D HT Schrittwieser J Antonoglou I Lai M Guez A et al Mastering chess and shogi by self play with a general reinforcement learning algorithm 2017 http arxiv org abs 1712 01815 accessed 8 8 19 4 Haenssle HA Fink C Schneiderbauer R et al Man against machine diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists Ann Oncol 2018 29 1836 42 5 Rajpurkar PIJ Zhu K Yang B et al CheXNet radiologist level pneumonia detection on chest X rays with deep learning CoRR abs 171105225 2017 6 Topol EJ High performancemedicine theconvergence of human andartifi cialintel ligence Nat Med 2019 25 44 56 7 Korosec K Waymo takes the wheel Tech Crunch 30 waymo takes the wheel self driving cars go fully driverless on california roads Accessed 8 2 19 8 Winkler Schwartz A Bissonnette V Mirchi N et al Artifi cial intelligence in medical education best practices using machine learning to assess surgical expertise in vir tual reality simulation J Surg Educ 2019 76 6 1681 90 9 Winkler Schwartz A Yilmaz R MirchiN etal Machinelearningidentifi cation ofsur gical and operative factors associated with surgical expertise in virtual reality simu lation JAMA Netw Open 2019 2 e198363 10 Druss BG Marcus SC Growth anddecentralizat